S²A-DL-EGBNet: Land Cover classification using an Attention-enabled Distributed Learning-based Encoder Generative Bidirectional Network

Land Cover (LC) classification is an effective technique that categorizes the Earth's surface into urban, forest, and agricultural land classes by utilizing remote sensing data. Nevertheless, the hyperspectral remote sensing images are afflicted with a lack of labeled data, spectral variability, and the curse of dimensionality, which limit their competence in remote sensing applications. Hence, this research presents a Spatial Split Attention-enabled Distributed Learning-based Encoder Generative Bidirectional Network (S²A-DL-EGBNet) model for accurate LC classification using hyperspectral images (HSIs). The model integrates a distributed learning module to process large datasets, and training is done in parallel by reducing complexity while improving scalability. Also, a Generative Adversarial Network (GAN)-based data balancing is employed to address the class imbalance problem and enables the model's effectiveness. Thereafter, the parallel Bidirectional Long Short-Term Memory (BiLSTM) is integrated to speed up the training process and minimize computation time. The research applies multiple feature extraction techniques, capturing complex features, scaling variations in photographic distortions, and illumination changes from the aspects of input data. Notably, the S²A-DL-EGBNet model is validated using the Hyperspectral Remote Sensing Scenes dataset, and the S²A-DL-EGBNet model shows remarkable performance by achieving 98.44% sensitivity, 98.84% accuracy, and 99.24% sensitivity on the Indian Pines dataset for training data of 90%.